1d38ceaf9SAlex Deucher /* 2d38ceaf9SAlex Deucher * Copyright 2009 Jerome Glisse. 3d38ceaf9SAlex Deucher * All Rights Reserved. 4d38ceaf9SAlex Deucher * 5d38ceaf9SAlex Deucher * Permission is hereby granted, free of charge, to any person obtaining a 6d38ceaf9SAlex Deucher * copy of this software and associated documentation files (the 7d38ceaf9SAlex Deucher * "Software"), to deal in the Software without restriction, including 8d38ceaf9SAlex Deucher * without limitation the rights to use, copy, modify, merge, publish, 9d38ceaf9SAlex Deucher * distribute, sub license, and/or sell copies of the Software, and to 10d38ceaf9SAlex Deucher * permit persons to whom the Software is furnished to do so, subject to 11d38ceaf9SAlex Deucher * the following conditions: 12d38ceaf9SAlex Deucher * 13d38ceaf9SAlex Deucher * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14d38ceaf9SAlex Deucher * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15d38ceaf9SAlex Deucher * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 16d38ceaf9SAlex Deucher * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 17d38ceaf9SAlex Deucher * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 18d38ceaf9SAlex Deucher * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 19d38ceaf9SAlex Deucher * USE OR OTHER DEALINGS IN THE SOFTWARE. 20d38ceaf9SAlex Deucher * 21d38ceaf9SAlex Deucher * The above copyright notice and this permission notice (including the 22d38ceaf9SAlex Deucher * next paragraph) shall be included in all copies or substantial portions 23d38ceaf9SAlex Deucher * of the Software. 24d38ceaf9SAlex Deucher * 25d38ceaf9SAlex Deucher */ 26d38ceaf9SAlex Deucher /* 27d38ceaf9SAlex Deucher * Authors: 28d38ceaf9SAlex Deucher * Jerome Glisse <glisse@freedesktop.org> 29d38ceaf9SAlex Deucher * Dave Airlie 30d38ceaf9SAlex Deucher */ 31d38ceaf9SAlex Deucher #include <linux/seq_file.h> 32d38ceaf9SAlex Deucher #include <linux/atomic.h> 33d38ceaf9SAlex Deucher #include <linux/wait.h> 34d38ceaf9SAlex Deucher #include <linux/kref.h> 35d38ceaf9SAlex Deucher #include <linux/slab.h> 36d38ceaf9SAlex Deucher #include <linux/firmware.h> 37d38ceaf9SAlex Deucher #include <drm/drmP.h> 38d38ceaf9SAlex Deucher #include "amdgpu.h" 39d38ceaf9SAlex Deucher #include "amdgpu_trace.h" 40d38ceaf9SAlex Deucher 41d38ceaf9SAlex Deucher /* 42d38ceaf9SAlex Deucher * Fences 43d38ceaf9SAlex Deucher * Fences mark an event in the GPUs pipeline and are used 44d38ceaf9SAlex Deucher * for GPU/CPU synchronization. When the fence is written, 45d38ceaf9SAlex Deucher * it is expected that all buffers associated with that fence 46d38ceaf9SAlex Deucher * are no longer in use by the associated ring on the GPU and 47d38ceaf9SAlex Deucher * that the the relevant GPU caches have been flushed. 48d38ceaf9SAlex Deucher */ 49d38ceaf9SAlex Deucher 50d38ceaf9SAlex Deucher /** 51d38ceaf9SAlex Deucher * amdgpu_fence_write - write a fence value 52d38ceaf9SAlex Deucher * 53d38ceaf9SAlex Deucher * @ring: ring the fence is associated with 54d38ceaf9SAlex Deucher * @seq: sequence number to write 55d38ceaf9SAlex Deucher * 56d38ceaf9SAlex Deucher * Writes a fence value to memory (all asics). 57d38ceaf9SAlex Deucher */ 58d38ceaf9SAlex Deucher static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq) 59d38ceaf9SAlex Deucher { 60d38ceaf9SAlex Deucher struct amdgpu_fence_driver *drv = &ring->fence_drv; 61d38ceaf9SAlex Deucher 62d38ceaf9SAlex Deucher if (drv->cpu_addr) 63d38ceaf9SAlex Deucher *drv->cpu_addr = cpu_to_le32(seq); 64d38ceaf9SAlex Deucher } 65d38ceaf9SAlex Deucher 66d38ceaf9SAlex Deucher /** 67d38ceaf9SAlex Deucher * amdgpu_fence_read - read a fence value 68d38ceaf9SAlex Deucher * 69d38ceaf9SAlex Deucher * @ring: ring the fence is associated with 70d38ceaf9SAlex Deucher * 71d38ceaf9SAlex Deucher * Reads a fence value from memory (all asics). 72d38ceaf9SAlex Deucher * Returns the value of the fence read from memory. 73d38ceaf9SAlex Deucher */ 74d38ceaf9SAlex Deucher static u32 amdgpu_fence_read(struct amdgpu_ring *ring) 75d38ceaf9SAlex Deucher { 76d38ceaf9SAlex Deucher struct amdgpu_fence_driver *drv = &ring->fence_drv; 77d38ceaf9SAlex Deucher u32 seq = 0; 78d38ceaf9SAlex Deucher 79d38ceaf9SAlex Deucher if (drv->cpu_addr) 80d38ceaf9SAlex Deucher seq = le32_to_cpu(*drv->cpu_addr); 81d38ceaf9SAlex Deucher else 82d38ceaf9SAlex Deucher seq = lower_32_bits(atomic64_read(&drv->last_seq)); 83d38ceaf9SAlex Deucher 84d38ceaf9SAlex Deucher return seq; 85d38ceaf9SAlex Deucher } 86d38ceaf9SAlex Deucher 87d38ceaf9SAlex Deucher /** 88d38ceaf9SAlex Deucher * amdgpu_fence_schedule_check - schedule lockup check 89d38ceaf9SAlex Deucher * 90d38ceaf9SAlex Deucher * @ring: pointer to struct amdgpu_ring 91d38ceaf9SAlex Deucher * 92d38ceaf9SAlex Deucher * Queues a delayed work item to check for lockups. 93d38ceaf9SAlex Deucher */ 94d38ceaf9SAlex Deucher static void amdgpu_fence_schedule_check(struct amdgpu_ring *ring) 95d38ceaf9SAlex Deucher { 96d38ceaf9SAlex Deucher /* 97d38ceaf9SAlex Deucher * Do not reset the timer here with mod_delayed_work, 98d38ceaf9SAlex Deucher * this can livelock in an interaction with TTM delayed destroy. 99d38ceaf9SAlex Deucher */ 100d38ceaf9SAlex Deucher queue_delayed_work(system_power_efficient_wq, 101d38ceaf9SAlex Deucher &ring->fence_drv.lockup_work, 102d38ceaf9SAlex Deucher AMDGPU_FENCE_JIFFIES_TIMEOUT); 103d38ceaf9SAlex Deucher } 104d38ceaf9SAlex Deucher 105d38ceaf9SAlex Deucher /** 106d38ceaf9SAlex Deucher * amdgpu_fence_emit - emit a fence on the requested ring 107d38ceaf9SAlex Deucher * 108d38ceaf9SAlex Deucher * @ring: ring the fence is associated with 109d38ceaf9SAlex Deucher * @owner: creator of the fence 110d38ceaf9SAlex Deucher * @fence: amdgpu fence object 111d38ceaf9SAlex Deucher * 112d38ceaf9SAlex Deucher * Emits a fence command on the requested ring (all asics). 113d38ceaf9SAlex Deucher * Returns 0 on success, -ENOMEM on failure. 114d38ceaf9SAlex Deucher */ 115d38ceaf9SAlex Deucher int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, 116d38ceaf9SAlex Deucher struct amdgpu_fence **fence) 117d38ceaf9SAlex Deucher { 118d38ceaf9SAlex Deucher struct amdgpu_device *adev = ring->adev; 119d38ceaf9SAlex Deucher 120d38ceaf9SAlex Deucher /* we are protected by the ring emission mutex */ 121d38ceaf9SAlex Deucher *fence = kmalloc(sizeof(struct amdgpu_fence), GFP_KERNEL); 122d38ceaf9SAlex Deucher if ((*fence) == NULL) { 123d38ceaf9SAlex Deucher return -ENOMEM; 124d38ceaf9SAlex Deucher } 125d38ceaf9SAlex Deucher (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx]; 126d38ceaf9SAlex Deucher (*fence)->ring = ring; 127d38ceaf9SAlex Deucher (*fence)->owner = owner; 128d38ceaf9SAlex Deucher fence_init(&(*fence)->base, &amdgpu_fence_ops, 129d38ceaf9SAlex Deucher &adev->fence_queue.lock, adev->fence_context + ring->idx, 130d38ceaf9SAlex Deucher (*fence)->seq); 131890ee23fSChunming Zhou amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, 132890ee23fSChunming Zhou (*fence)->seq, 133890ee23fSChunming Zhou AMDGPU_FENCE_FLAG_INT); 134d38ceaf9SAlex Deucher trace_amdgpu_fence_emit(ring->adev->ddev, ring->idx, (*fence)->seq); 135d38ceaf9SAlex Deucher return 0; 136d38ceaf9SAlex Deucher } 137d38ceaf9SAlex Deucher 138d38ceaf9SAlex Deucher /** 139d38ceaf9SAlex Deucher * amdgpu_fence_check_signaled - callback from fence_queue 140d38ceaf9SAlex Deucher * 141d38ceaf9SAlex Deucher * this function is called with fence_queue lock held, which is also used 142d38ceaf9SAlex Deucher * for the fence locking itself, so unlocked variants are used for 143d38ceaf9SAlex Deucher * fence_signal, and remove_wait_queue. 144d38ceaf9SAlex Deucher */ 145d38ceaf9SAlex Deucher static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key) 146d38ceaf9SAlex Deucher { 147d38ceaf9SAlex Deucher struct amdgpu_fence *fence; 148d38ceaf9SAlex Deucher struct amdgpu_device *adev; 149d38ceaf9SAlex Deucher u64 seq; 150d38ceaf9SAlex Deucher int ret; 151d38ceaf9SAlex Deucher 152d38ceaf9SAlex Deucher fence = container_of(wait, struct amdgpu_fence, fence_wake); 153d38ceaf9SAlex Deucher adev = fence->ring->adev; 154d38ceaf9SAlex Deucher 155d38ceaf9SAlex Deucher /* 156d38ceaf9SAlex Deucher * We cannot use amdgpu_fence_process here because we're already 157d38ceaf9SAlex Deucher * in the waitqueue, in a call from wake_up_all. 158d38ceaf9SAlex Deucher */ 159d38ceaf9SAlex Deucher seq = atomic64_read(&fence->ring->fence_drv.last_seq); 160d38ceaf9SAlex Deucher if (seq >= fence->seq) { 161d38ceaf9SAlex Deucher ret = fence_signal_locked(&fence->base); 162d38ceaf9SAlex Deucher if (!ret) 163d38ceaf9SAlex Deucher FENCE_TRACE(&fence->base, "signaled from irq context\n"); 164d38ceaf9SAlex Deucher else 165d38ceaf9SAlex Deucher FENCE_TRACE(&fence->base, "was already signaled\n"); 166d38ceaf9SAlex Deucher 167d38ceaf9SAlex Deucher __remove_wait_queue(&adev->fence_queue, &fence->fence_wake); 168d38ceaf9SAlex Deucher fence_put(&fence->base); 169d38ceaf9SAlex Deucher } else 170d38ceaf9SAlex Deucher FENCE_TRACE(&fence->base, "pending\n"); 171d38ceaf9SAlex Deucher return 0; 172d38ceaf9SAlex Deucher } 173d38ceaf9SAlex Deucher 174d38ceaf9SAlex Deucher /** 175d38ceaf9SAlex Deucher * amdgpu_fence_activity - check for fence activity 176d38ceaf9SAlex Deucher * 177d38ceaf9SAlex Deucher * @ring: pointer to struct amdgpu_ring 178d38ceaf9SAlex Deucher * 179d38ceaf9SAlex Deucher * Checks the current fence value and calculates the last 180d38ceaf9SAlex Deucher * signalled fence value. Returns true if activity occured 181d38ceaf9SAlex Deucher * on the ring, and the fence_queue should be waken up. 182d38ceaf9SAlex Deucher */ 183d38ceaf9SAlex Deucher static bool amdgpu_fence_activity(struct amdgpu_ring *ring) 184d38ceaf9SAlex Deucher { 185d38ceaf9SAlex Deucher uint64_t seq, last_seq, last_emitted; 186d38ceaf9SAlex Deucher unsigned count_loop = 0; 187d38ceaf9SAlex Deucher bool wake = false; 188d38ceaf9SAlex Deucher 189d38ceaf9SAlex Deucher /* Note there is a scenario here for an infinite loop but it's 190d38ceaf9SAlex Deucher * very unlikely to happen. For it to happen, the current polling 191d38ceaf9SAlex Deucher * process need to be interrupted by another process and another 192d38ceaf9SAlex Deucher * process needs to update the last_seq btw the atomic read and 193d38ceaf9SAlex Deucher * xchg of the current process. 194d38ceaf9SAlex Deucher * 195d38ceaf9SAlex Deucher * More over for this to go in infinite loop there need to be 19686c2b790SJammy Zhou * continuously new fence signaled ie amdgpu_fence_read needs 197d38ceaf9SAlex Deucher * to return a different value each time for both the currently 198d38ceaf9SAlex Deucher * polling process and the other process that xchg the last_seq 199d38ceaf9SAlex Deucher * btw atomic read and xchg of the current process. And the 200d38ceaf9SAlex Deucher * value the other process set as last seq must be higher than 201d38ceaf9SAlex Deucher * the seq value we just read. Which means that current process 20286c2b790SJammy Zhou * need to be interrupted after amdgpu_fence_read and before 203d38ceaf9SAlex Deucher * atomic xchg. 204d38ceaf9SAlex Deucher * 205d38ceaf9SAlex Deucher * To be even more safe we count the number of time we loop and 206d38ceaf9SAlex Deucher * we bail after 10 loop just accepting the fact that we might 207d38ceaf9SAlex Deucher * have temporarly set the last_seq not to the true real last 208d38ceaf9SAlex Deucher * seq but to an older one. 209d38ceaf9SAlex Deucher */ 210d38ceaf9SAlex Deucher last_seq = atomic64_read(&ring->fence_drv.last_seq); 211d38ceaf9SAlex Deucher do { 212d38ceaf9SAlex Deucher last_emitted = ring->fence_drv.sync_seq[ring->idx]; 213d38ceaf9SAlex Deucher seq = amdgpu_fence_read(ring); 214d38ceaf9SAlex Deucher seq |= last_seq & 0xffffffff00000000LL; 215d38ceaf9SAlex Deucher if (seq < last_seq) { 216d38ceaf9SAlex Deucher seq &= 0xffffffff; 217d38ceaf9SAlex Deucher seq |= last_emitted & 0xffffffff00000000LL; 218d38ceaf9SAlex Deucher } 219d38ceaf9SAlex Deucher 220d38ceaf9SAlex Deucher if (seq <= last_seq || seq > last_emitted) { 221d38ceaf9SAlex Deucher break; 222d38ceaf9SAlex Deucher } 223d38ceaf9SAlex Deucher /* If we loop over we don't want to return without 224d38ceaf9SAlex Deucher * checking if a fence is signaled as it means that the 225d38ceaf9SAlex Deucher * seq we just read is different from the previous on. 226d38ceaf9SAlex Deucher */ 227d38ceaf9SAlex Deucher wake = true; 228d38ceaf9SAlex Deucher last_seq = seq; 229d38ceaf9SAlex Deucher if ((count_loop++) > 10) { 230d38ceaf9SAlex Deucher /* We looped over too many time leave with the 231d38ceaf9SAlex Deucher * fact that we might have set an older fence 232d38ceaf9SAlex Deucher * seq then the current real last seq as signaled 233d38ceaf9SAlex Deucher * by the hw. 234d38ceaf9SAlex Deucher */ 235d38ceaf9SAlex Deucher break; 236d38ceaf9SAlex Deucher } 237d38ceaf9SAlex Deucher } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); 238d38ceaf9SAlex Deucher 239d38ceaf9SAlex Deucher if (seq < last_emitted) 240d38ceaf9SAlex Deucher amdgpu_fence_schedule_check(ring); 241d38ceaf9SAlex Deucher 242d38ceaf9SAlex Deucher return wake; 243d38ceaf9SAlex Deucher } 244d38ceaf9SAlex Deucher 245d38ceaf9SAlex Deucher /** 246d38ceaf9SAlex Deucher * amdgpu_fence_check_lockup - check for hardware lockup 247d38ceaf9SAlex Deucher * 248d38ceaf9SAlex Deucher * @work: delayed work item 249d38ceaf9SAlex Deucher * 250d38ceaf9SAlex Deucher * Checks for fence activity and if there is none probe 251d38ceaf9SAlex Deucher * the hardware if a lockup occured. 252d38ceaf9SAlex Deucher */ 253d38ceaf9SAlex Deucher static void amdgpu_fence_check_lockup(struct work_struct *work) 254d38ceaf9SAlex Deucher { 255d38ceaf9SAlex Deucher struct amdgpu_fence_driver *fence_drv; 256d38ceaf9SAlex Deucher struct amdgpu_ring *ring; 257d38ceaf9SAlex Deucher 258d38ceaf9SAlex Deucher fence_drv = container_of(work, struct amdgpu_fence_driver, 259d38ceaf9SAlex Deucher lockup_work.work); 260d38ceaf9SAlex Deucher ring = fence_drv->ring; 261d38ceaf9SAlex Deucher 262d38ceaf9SAlex Deucher if (!down_read_trylock(&ring->adev->exclusive_lock)) { 263d38ceaf9SAlex Deucher /* just reschedule the check if a reset is going on */ 264d38ceaf9SAlex Deucher amdgpu_fence_schedule_check(ring); 265d38ceaf9SAlex Deucher return; 266d38ceaf9SAlex Deucher } 267d38ceaf9SAlex Deucher 268d38ceaf9SAlex Deucher if (amdgpu_fence_activity(ring)) 269d38ceaf9SAlex Deucher wake_up_all(&ring->adev->fence_queue); 270d38ceaf9SAlex Deucher else if (amdgpu_ring_is_lockup(ring)) { 271d38ceaf9SAlex Deucher /* good news we believe it's a lockup */ 272d38ceaf9SAlex Deucher dev_warn(ring->adev->dev, "GPU lockup (current fence id " 273d38ceaf9SAlex Deucher "0x%016llx last fence id 0x%016llx on ring %d)\n", 274d38ceaf9SAlex Deucher (uint64_t)atomic64_read(&fence_drv->last_seq), 275d38ceaf9SAlex Deucher fence_drv->sync_seq[ring->idx], ring->idx); 276d38ceaf9SAlex Deucher 277d38ceaf9SAlex Deucher /* remember that we need an reset */ 278d38ceaf9SAlex Deucher ring->adev->needs_reset = true; 279d38ceaf9SAlex Deucher wake_up_all(&ring->adev->fence_queue); 280d38ceaf9SAlex Deucher } 281d38ceaf9SAlex Deucher up_read(&ring->adev->exclusive_lock); 282d38ceaf9SAlex Deucher } 283d38ceaf9SAlex Deucher 284d38ceaf9SAlex Deucher /** 285d38ceaf9SAlex Deucher * amdgpu_fence_process - process a fence 286d38ceaf9SAlex Deucher * 287d38ceaf9SAlex Deucher * @adev: amdgpu_device pointer 288d38ceaf9SAlex Deucher * @ring: ring index the fence is associated with 289d38ceaf9SAlex Deucher * 290d38ceaf9SAlex Deucher * Checks the current fence value and wakes the fence queue 291d38ceaf9SAlex Deucher * if the sequence number has increased (all asics). 292d38ceaf9SAlex Deucher */ 293d38ceaf9SAlex Deucher void amdgpu_fence_process(struct amdgpu_ring *ring) 294d38ceaf9SAlex Deucher { 295d38ceaf9SAlex Deucher uint64_t seq, last_seq, last_emitted; 296d38ceaf9SAlex Deucher unsigned count_loop = 0; 297d38ceaf9SAlex Deucher bool wake = false; 298176e1ab1SChunming Zhou unsigned long irqflags; 299d38ceaf9SAlex Deucher 300d38ceaf9SAlex Deucher /* Note there is a scenario here for an infinite loop but it's 301d38ceaf9SAlex Deucher * very unlikely to happen. For it to happen, the current polling 302d38ceaf9SAlex Deucher * process need to be interrupted by another process and another 303d38ceaf9SAlex Deucher * process needs to update the last_seq btw the atomic read and 304d38ceaf9SAlex Deucher * xchg of the current process. 305d38ceaf9SAlex Deucher * 306d38ceaf9SAlex Deucher * More over for this to go in infinite loop there need to be 307d38ceaf9SAlex Deucher * continuously new fence signaled ie amdgpu_fence_read needs 308d38ceaf9SAlex Deucher * to return a different value each time for both the currently 309d38ceaf9SAlex Deucher * polling process and the other process that xchg the last_seq 310d38ceaf9SAlex Deucher * btw atomic read and xchg of the current process. And the 311d38ceaf9SAlex Deucher * value the other process set as last seq must be higher than 312d38ceaf9SAlex Deucher * the seq value we just read. Which means that current process 313d38ceaf9SAlex Deucher * need to be interrupted after amdgpu_fence_read and before 314d38ceaf9SAlex Deucher * atomic xchg. 315d38ceaf9SAlex Deucher * 316d38ceaf9SAlex Deucher * To be even more safe we count the number of time we loop and 317d38ceaf9SAlex Deucher * we bail after 10 loop just accepting the fact that we might 318d38ceaf9SAlex Deucher * have temporarly set the last_seq not to the true real last 319d38ceaf9SAlex Deucher * seq but to an older one. 320d38ceaf9SAlex Deucher */ 321176e1ab1SChunming Zhou spin_lock_irqsave(&ring->fence_lock, irqflags); 322d38ceaf9SAlex Deucher last_seq = atomic64_read(&ring->fence_drv.last_seq); 323d38ceaf9SAlex Deucher do { 324d38ceaf9SAlex Deucher last_emitted = ring->fence_drv.sync_seq[ring->idx]; 325d38ceaf9SAlex Deucher seq = amdgpu_fence_read(ring); 326d38ceaf9SAlex Deucher seq |= last_seq & 0xffffffff00000000LL; 327d38ceaf9SAlex Deucher if (seq < last_seq) { 328d38ceaf9SAlex Deucher seq &= 0xffffffff; 329d38ceaf9SAlex Deucher seq |= last_emitted & 0xffffffff00000000LL; 330d38ceaf9SAlex Deucher } 331d38ceaf9SAlex Deucher 332d38ceaf9SAlex Deucher if (seq <= last_seq || seq > last_emitted) { 333d38ceaf9SAlex Deucher break; 334d38ceaf9SAlex Deucher } 335d38ceaf9SAlex Deucher /* If we loop over we don't want to return without 336d38ceaf9SAlex Deucher * checking if a fence is signaled as it means that the 337d38ceaf9SAlex Deucher * seq we just read is different from the previous on. 338d38ceaf9SAlex Deucher */ 339d38ceaf9SAlex Deucher wake = true; 340d38ceaf9SAlex Deucher last_seq = seq; 341d38ceaf9SAlex Deucher if ((count_loop++) > 10) { 342d38ceaf9SAlex Deucher /* We looped over too many time leave with the 343d38ceaf9SAlex Deucher * fact that we might have set an older fence 344d38ceaf9SAlex Deucher * seq then the current real last seq as signaled 345d38ceaf9SAlex Deucher * by the hw. 346d38ceaf9SAlex Deucher */ 347d38ceaf9SAlex Deucher break; 348d38ceaf9SAlex Deucher } 349d38ceaf9SAlex Deucher } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); 350d38ceaf9SAlex Deucher 351e0d8f3c3SChunming Zhou if (wake) { 352e0d8f3c3SChunming Zhou if (amdgpu_enable_scheduler) { 353e0d8f3c3SChunming Zhou uint64_t handled_seq = 354e0d8f3c3SChunming Zhou amd_sched_get_handled_seq(ring->scheduler); 355e0d8f3c3SChunming Zhou uint64_t latest_seq = 356e0d8f3c3SChunming Zhou atomic64_read(&ring->fence_drv.last_seq); 357e0d8f3c3SChunming Zhou if (handled_seq == latest_seq) { 358e0d8f3c3SChunming Zhou DRM_ERROR("ring %d, EOP without seq update (lastest_seq=%llu)\n", 359e0d8f3c3SChunming Zhou ring->idx, latest_seq); 360176e1ab1SChunming Zhou goto exit; 361e0d8f3c3SChunming Zhou } 362e0d8f3c3SChunming Zhou do { 363e0d8f3c3SChunming Zhou amd_sched_isr(ring->scheduler); 364e0d8f3c3SChunming Zhou } while (amd_sched_get_handled_seq(ring->scheduler) < latest_seq); 365e0d8f3c3SChunming Zhou } 366e0d8f3c3SChunming Zhou 367d38ceaf9SAlex Deucher wake_up_all(&ring->adev->fence_queue); 368d38ceaf9SAlex Deucher } 369176e1ab1SChunming Zhou exit: 370176e1ab1SChunming Zhou spin_unlock_irqrestore(&ring->fence_lock, irqflags); 371e0d8f3c3SChunming Zhou } 372d38ceaf9SAlex Deucher 373d38ceaf9SAlex Deucher /** 374d38ceaf9SAlex Deucher * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled 375d38ceaf9SAlex Deucher * 376d38ceaf9SAlex Deucher * @ring: ring the fence is associated with 377d38ceaf9SAlex Deucher * @seq: sequence number 378d38ceaf9SAlex Deucher * 379d38ceaf9SAlex Deucher * Check if the last signaled fence sequnce number is >= the requested 380d38ceaf9SAlex Deucher * sequence number (all asics). 381d38ceaf9SAlex Deucher * Returns true if the fence has signaled (current fence value 382d38ceaf9SAlex Deucher * is >= requested value) or false if it has not (current fence 383d38ceaf9SAlex Deucher * value is < the requested value. Helper function for 384d38ceaf9SAlex Deucher * amdgpu_fence_signaled(). 385d38ceaf9SAlex Deucher */ 386d38ceaf9SAlex Deucher static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq) 387d38ceaf9SAlex Deucher { 388d38ceaf9SAlex Deucher if (atomic64_read(&ring->fence_drv.last_seq) >= seq) 389d38ceaf9SAlex Deucher return true; 390d38ceaf9SAlex Deucher 391d38ceaf9SAlex Deucher /* poll new last sequence at least once */ 392d38ceaf9SAlex Deucher amdgpu_fence_process(ring); 393d38ceaf9SAlex Deucher if (atomic64_read(&ring->fence_drv.last_seq) >= seq) 394d38ceaf9SAlex Deucher return true; 395d38ceaf9SAlex Deucher 396d38ceaf9SAlex Deucher return false; 397d38ceaf9SAlex Deucher } 398d38ceaf9SAlex Deucher 399d38ceaf9SAlex Deucher static bool amdgpu_fence_is_signaled(struct fence *f) 400d38ceaf9SAlex Deucher { 401d38ceaf9SAlex Deucher struct amdgpu_fence *fence = to_amdgpu_fence(f); 402d38ceaf9SAlex Deucher struct amdgpu_ring *ring = fence->ring; 403d38ceaf9SAlex Deucher struct amdgpu_device *adev = ring->adev; 404d38ceaf9SAlex Deucher 405d38ceaf9SAlex Deucher if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) 406d38ceaf9SAlex Deucher return true; 407d38ceaf9SAlex Deucher 408d38ceaf9SAlex Deucher if (down_read_trylock(&adev->exclusive_lock)) { 409d38ceaf9SAlex Deucher amdgpu_fence_process(ring); 410d38ceaf9SAlex Deucher up_read(&adev->exclusive_lock); 411d38ceaf9SAlex Deucher 412d38ceaf9SAlex Deucher if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) 413d38ceaf9SAlex Deucher return true; 414d38ceaf9SAlex Deucher } 415d38ceaf9SAlex Deucher return false; 416d38ceaf9SAlex Deucher } 417d38ceaf9SAlex Deucher 418d38ceaf9SAlex Deucher /** 419d38ceaf9SAlex Deucher * amdgpu_fence_enable_signaling - enable signalling on fence 420d38ceaf9SAlex Deucher * @fence: fence 421d38ceaf9SAlex Deucher * 422d38ceaf9SAlex Deucher * This function is called with fence_queue lock held, and adds a callback 423d38ceaf9SAlex Deucher * to fence_queue that checks if this fence is signaled, and if so it 424d38ceaf9SAlex Deucher * signals the fence and removes itself. 425d38ceaf9SAlex Deucher */ 426d38ceaf9SAlex Deucher static bool amdgpu_fence_enable_signaling(struct fence *f) 427d38ceaf9SAlex Deucher { 428d38ceaf9SAlex Deucher struct amdgpu_fence *fence = to_amdgpu_fence(f); 429d38ceaf9SAlex Deucher struct amdgpu_ring *ring = fence->ring; 430d38ceaf9SAlex Deucher struct amdgpu_device *adev = ring->adev; 431d38ceaf9SAlex Deucher 432d38ceaf9SAlex Deucher if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) 433d38ceaf9SAlex Deucher return false; 434d38ceaf9SAlex Deucher 435d38ceaf9SAlex Deucher fence->fence_wake.flags = 0; 436d38ceaf9SAlex Deucher fence->fence_wake.private = NULL; 437d38ceaf9SAlex Deucher fence->fence_wake.func = amdgpu_fence_check_signaled; 438d38ceaf9SAlex Deucher __add_wait_queue(&adev->fence_queue, &fence->fence_wake); 439d38ceaf9SAlex Deucher fence_get(f); 440d38ceaf9SAlex Deucher FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx); 441d38ceaf9SAlex Deucher return true; 442d38ceaf9SAlex Deucher } 443d38ceaf9SAlex Deucher 444d38ceaf9SAlex Deucher /** 445d38ceaf9SAlex Deucher * amdgpu_fence_signaled - check if a fence has signaled 446d38ceaf9SAlex Deucher * 447d38ceaf9SAlex Deucher * @fence: amdgpu fence object 448d38ceaf9SAlex Deucher * 449d38ceaf9SAlex Deucher * Check if the requested fence has signaled (all asics). 450d38ceaf9SAlex Deucher * Returns true if the fence has signaled or false if it has not. 451d38ceaf9SAlex Deucher */ 452d38ceaf9SAlex Deucher bool amdgpu_fence_signaled(struct amdgpu_fence *fence) 453d38ceaf9SAlex Deucher { 454d38ceaf9SAlex Deucher if (!fence) 455d38ceaf9SAlex Deucher return true; 456d38ceaf9SAlex Deucher 457d38ceaf9SAlex Deucher if (amdgpu_fence_seq_signaled(fence->ring, fence->seq)) { 458d38ceaf9SAlex Deucher if (!fence_signal(&fence->base)) 459d38ceaf9SAlex Deucher FENCE_TRACE(&fence->base, "signaled from amdgpu_fence_signaled\n"); 460d38ceaf9SAlex Deucher return true; 461d38ceaf9SAlex Deucher } 462d38ceaf9SAlex Deucher 463d38ceaf9SAlex Deucher return false; 464d38ceaf9SAlex Deucher } 465d38ceaf9SAlex Deucher 466d38ceaf9SAlex Deucher /** 467d38ceaf9SAlex Deucher * amdgpu_fence_any_seq_signaled - check if any sequence number is signaled 468d38ceaf9SAlex Deucher * 469d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 470d38ceaf9SAlex Deucher * @seq: sequence numbers 471d38ceaf9SAlex Deucher * 472d38ceaf9SAlex Deucher * Check if the last signaled fence sequnce number is >= the requested 473d38ceaf9SAlex Deucher * sequence number (all asics). 474d38ceaf9SAlex Deucher * Returns true if any has signaled (current value is >= requested value) 475d38ceaf9SAlex Deucher * or false if it has not. Helper function for amdgpu_fence_wait_seq. 476d38ceaf9SAlex Deucher */ 477d38ceaf9SAlex Deucher static bool amdgpu_fence_any_seq_signaled(struct amdgpu_device *adev, u64 *seq) 478d38ceaf9SAlex Deucher { 479d38ceaf9SAlex Deucher unsigned i; 480d38ceaf9SAlex Deucher 481d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 482d38ceaf9SAlex Deucher if (!adev->rings[i] || !seq[i]) 483d38ceaf9SAlex Deucher continue; 484d38ceaf9SAlex Deucher 485d38ceaf9SAlex Deucher if (amdgpu_fence_seq_signaled(adev->rings[i], seq[i])) 486d38ceaf9SAlex Deucher return true; 487d38ceaf9SAlex Deucher } 488d38ceaf9SAlex Deucher 489d38ceaf9SAlex Deucher return false; 490d38ceaf9SAlex Deucher } 491d38ceaf9SAlex Deucher 492d38ceaf9SAlex Deucher /** 493d38ceaf9SAlex Deucher * amdgpu_fence_wait_seq_timeout - wait for a specific sequence numbers 494d38ceaf9SAlex Deucher * 495d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 496d38ceaf9SAlex Deucher * @target_seq: sequence number(s) we want to wait for 497d38ceaf9SAlex Deucher * @intr: use interruptable sleep 498d38ceaf9SAlex Deucher * @timeout: maximum time to wait, or MAX_SCHEDULE_TIMEOUT for infinite wait 499d38ceaf9SAlex Deucher * 500d38ceaf9SAlex Deucher * Wait for the requested sequence number(s) to be written by any ring 501d38ceaf9SAlex Deucher * (all asics). Sequnce number array is indexed by ring id. 502d38ceaf9SAlex Deucher * @intr selects whether to use interruptable (true) or non-interruptable 503d38ceaf9SAlex Deucher * (false) sleep when waiting for the sequence number. Helper function 504d38ceaf9SAlex Deucher * for amdgpu_fence_wait_*(). 505d38ceaf9SAlex Deucher * Returns remaining time if the sequence number has passed, 0 when 506d38ceaf9SAlex Deucher * the wait timeout, or an error for all other cases. 507d38ceaf9SAlex Deucher * -EDEADLK is returned when a GPU lockup has been detected. 508d38ceaf9SAlex Deucher */ 50903507c4fSChristian König static long amdgpu_fence_wait_seq_timeout(struct amdgpu_device *adev, 51003507c4fSChristian König u64 *target_seq, bool intr, 51103507c4fSChristian König long timeout) 512d38ceaf9SAlex Deucher { 513d38ceaf9SAlex Deucher uint64_t last_seq[AMDGPU_MAX_RINGS]; 514d38ceaf9SAlex Deucher bool signaled; 515332300b9Smonk.liu int i; 516332300b9Smonk.liu long r; 517d38ceaf9SAlex Deucher 51825f45e63SJack Xiao if (timeout == 0) { 51925f45e63SJack Xiao return amdgpu_fence_any_seq_signaled(adev, target_seq); 52025f45e63SJack Xiao } 52125f45e63SJack Xiao 522d38ceaf9SAlex Deucher while (!amdgpu_fence_any_seq_signaled(adev, target_seq)) { 523d38ceaf9SAlex Deucher 524d38ceaf9SAlex Deucher /* Save current sequence values, used to check for GPU lockups */ 525d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 526d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 527d38ceaf9SAlex Deucher 528d38ceaf9SAlex Deucher if (!ring || !target_seq[i]) 529d38ceaf9SAlex Deucher continue; 530d38ceaf9SAlex Deucher 531d38ceaf9SAlex Deucher last_seq[i] = atomic64_read(&ring->fence_drv.last_seq); 532d38ceaf9SAlex Deucher trace_amdgpu_fence_wait_begin(adev->ddev, i, target_seq[i]); 533d38ceaf9SAlex Deucher } 534d38ceaf9SAlex Deucher 535d38ceaf9SAlex Deucher if (intr) { 536d38ceaf9SAlex Deucher r = wait_event_interruptible_timeout(adev->fence_queue, ( 537d38ceaf9SAlex Deucher (signaled = amdgpu_fence_any_seq_signaled(adev, target_seq)) 538d38ceaf9SAlex Deucher || adev->needs_reset), AMDGPU_FENCE_JIFFIES_TIMEOUT); 539d38ceaf9SAlex Deucher } else { 540d38ceaf9SAlex Deucher r = wait_event_timeout(adev->fence_queue, ( 541d38ceaf9SAlex Deucher (signaled = amdgpu_fence_any_seq_signaled(adev, target_seq)) 542d38ceaf9SAlex Deucher || adev->needs_reset), AMDGPU_FENCE_JIFFIES_TIMEOUT); 543d38ceaf9SAlex Deucher } 544d38ceaf9SAlex Deucher 545d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 546d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 547d38ceaf9SAlex Deucher 548d38ceaf9SAlex Deucher if (!ring || !target_seq[i]) 549d38ceaf9SAlex Deucher continue; 550d38ceaf9SAlex Deucher 551d38ceaf9SAlex Deucher trace_amdgpu_fence_wait_end(adev->ddev, i, target_seq[i]); 552d38ceaf9SAlex Deucher } 553d38ceaf9SAlex Deucher 554d38ceaf9SAlex Deucher if (unlikely(r < 0)) 555d38ceaf9SAlex Deucher return r; 556d38ceaf9SAlex Deucher 557d38ceaf9SAlex Deucher if (unlikely(!signaled)) { 558d38ceaf9SAlex Deucher 559d38ceaf9SAlex Deucher if (adev->needs_reset) 560d38ceaf9SAlex Deucher return -EDEADLK; 561d38ceaf9SAlex Deucher 562d38ceaf9SAlex Deucher /* we were interrupted for some reason and fence 563d38ceaf9SAlex Deucher * isn't signaled yet, resume waiting */ 564d38ceaf9SAlex Deucher if (r) 565d38ceaf9SAlex Deucher continue; 566d38ceaf9SAlex Deucher 567d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 568d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 569d38ceaf9SAlex Deucher 570d38ceaf9SAlex Deucher if (!ring || !target_seq[i]) 571d38ceaf9SAlex Deucher continue; 572d38ceaf9SAlex Deucher 573d38ceaf9SAlex Deucher if (last_seq[i] != atomic64_read(&ring->fence_drv.last_seq)) 574d38ceaf9SAlex Deucher break; 575d38ceaf9SAlex Deucher } 576d38ceaf9SAlex Deucher 577d38ceaf9SAlex Deucher if (i != AMDGPU_MAX_RINGS) 578d38ceaf9SAlex Deucher continue; 579d38ceaf9SAlex Deucher 580d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 581d38ceaf9SAlex Deucher if (!adev->rings[i] || !target_seq[i]) 582d38ceaf9SAlex Deucher continue; 583d38ceaf9SAlex Deucher 584d38ceaf9SAlex Deucher if (amdgpu_ring_is_lockup(adev->rings[i])) 585d38ceaf9SAlex Deucher break; 586d38ceaf9SAlex Deucher } 587d38ceaf9SAlex Deucher 588d38ceaf9SAlex Deucher if (i < AMDGPU_MAX_RINGS) { 589d38ceaf9SAlex Deucher /* good news we believe it's a lockup */ 590d38ceaf9SAlex Deucher dev_warn(adev->dev, "GPU lockup (waiting for " 591d38ceaf9SAlex Deucher "0x%016llx last fence id 0x%016llx on" 592d38ceaf9SAlex Deucher " ring %d)\n", 593d38ceaf9SAlex Deucher target_seq[i], last_seq[i], i); 594d38ceaf9SAlex Deucher 595d38ceaf9SAlex Deucher /* remember that we need an reset */ 596d38ceaf9SAlex Deucher adev->needs_reset = true; 597d38ceaf9SAlex Deucher wake_up_all(&adev->fence_queue); 598d38ceaf9SAlex Deucher return -EDEADLK; 599d38ceaf9SAlex Deucher } 600d38ceaf9SAlex Deucher 601d38ceaf9SAlex Deucher if (timeout < MAX_SCHEDULE_TIMEOUT) { 602d38ceaf9SAlex Deucher timeout -= AMDGPU_FENCE_JIFFIES_TIMEOUT; 603d38ceaf9SAlex Deucher if (timeout <= 0) { 604d38ceaf9SAlex Deucher return 0; 605d38ceaf9SAlex Deucher } 606d38ceaf9SAlex Deucher } 607d38ceaf9SAlex Deucher } 608d38ceaf9SAlex Deucher } 609d38ceaf9SAlex Deucher return timeout; 610d38ceaf9SAlex Deucher } 611d38ceaf9SAlex Deucher 612d38ceaf9SAlex Deucher /** 613d38ceaf9SAlex Deucher * amdgpu_fence_wait - wait for a fence to signal 614d38ceaf9SAlex Deucher * 615d38ceaf9SAlex Deucher * @fence: amdgpu fence object 616d38ceaf9SAlex Deucher * @intr: use interruptable sleep 617d38ceaf9SAlex Deucher * 618d38ceaf9SAlex Deucher * Wait for the requested fence to signal (all asics). 619d38ceaf9SAlex Deucher * @intr selects whether to use interruptable (true) or non-interruptable 620d38ceaf9SAlex Deucher * (false) sleep when waiting for the fence. 621d38ceaf9SAlex Deucher * Returns 0 if the fence has passed, error for all other cases. 622d38ceaf9SAlex Deucher */ 623d38ceaf9SAlex Deucher int amdgpu_fence_wait(struct amdgpu_fence *fence, bool intr) 624d38ceaf9SAlex Deucher { 625d38ceaf9SAlex Deucher long r; 626d38ceaf9SAlex Deucher 6272e536084Smonk.liu r = fence_wait_timeout(&fence->base, intr, MAX_SCHEDULE_TIMEOUT); 6282e536084Smonk.liu if (r < 0) 629d38ceaf9SAlex Deucher return r; 630d38ceaf9SAlex Deucher return 0; 631d38ceaf9SAlex Deucher } 632d38ceaf9SAlex Deucher 633d38ceaf9SAlex Deucher /** 634d38ceaf9SAlex Deucher * amdgpu_fence_wait_any - wait for a fence to signal on any ring 635d38ceaf9SAlex Deucher * 636d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 637d38ceaf9SAlex Deucher * @fences: amdgpu fence object(s) 638d38ceaf9SAlex Deucher * @intr: use interruptable sleep 639d38ceaf9SAlex Deucher * 640d38ceaf9SAlex Deucher * Wait for any requested fence to signal (all asics). Fence 641d38ceaf9SAlex Deucher * array is indexed by ring id. @intr selects whether to use 642d38ceaf9SAlex Deucher * interruptable (true) or non-interruptable (false) sleep when 643d38ceaf9SAlex Deucher * waiting for the fences. Used by the suballocator. 644d38ceaf9SAlex Deucher * Returns 0 if any fence has passed, error for all other cases. 645d38ceaf9SAlex Deucher */ 646d38ceaf9SAlex Deucher int amdgpu_fence_wait_any(struct amdgpu_device *adev, 647d38ceaf9SAlex Deucher struct amdgpu_fence **fences, 648d38ceaf9SAlex Deucher bool intr) 649d38ceaf9SAlex Deucher { 650d38ceaf9SAlex Deucher uint64_t seq[AMDGPU_MAX_RINGS]; 651d38ceaf9SAlex Deucher unsigned i, num_rings = 0; 652d38ceaf9SAlex Deucher long r; 653d38ceaf9SAlex Deucher 654d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 655d38ceaf9SAlex Deucher seq[i] = 0; 656d38ceaf9SAlex Deucher 657d38ceaf9SAlex Deucher if (!fences[i]) { 658d38ceaf9SAlex Deucher continue; 659d38ceaf9SAlex Deucher } 660d38ceaf9SAlex Deucher 661d38ceaf9SAlex Deucher seq[i] = fences[i]->seq; 662d38ceaf9SAlex Deucher ++num_rings; 663d38ceaf9SAlex Deucher } 664d38ceaf9SAlex Deucher 665d38ceaf9SAlex Deucher /* nothing to wait for ? */ 666d38ceaf9SAlex Deucher if (num_rings == 0) 667d38ceaf9SAlex Deucher return -ENOENT; 668d38ceaf9SAlex Deucher 669d38ceaf9SAlex Deucher r = amdgpu_fence_wait_seq_timeout(adev, seq, intr, MAX_SCHEDULE_TIMEOUT); 670d38ceaf9SAlex Deucher if (r < 0) { 671d38ceaf9SAlex Deucher return r; 672d38ceaf9SAlex Deucher } 673d38ceaf9SAlex Deucher return 0; 674d38ceaf9SAlex Deucher } 675d38ceaf9SAlex Deucher 676d38ceaf9SAlex Deucher /** 677d38ceaf9SAlex Deucher * amdgpu_fence_wait_next - wait for the next fence to signal 678d38ceaf9SAlex Deucher * 679d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 680d38ceaf9SAlex Deucher * @ring: ring index the fence is associated with 681d38ceaf9SAlex Deucher * 682d38ceaf9SAlex Deucher * Wait for the next fence on the requested ring to signal (all asics). 683d38ceaf9SAlex Deucher * Returns 0 if the next fence has passed, error for all other cases. 684d38ceaf9SAlex Deucher * Caller must hold ring lock. 685d38ceaf9SAlex Deucher */ 686d38ceaf9SAlex Deucher int amdgpu_fence_wait_next(struct amdgpu_ring *ring) 687d38ceaf9SAlex Deucher { 688d38ceaf9SAlex Deucher uint64_t seq[AMDGPU_MAX_RINGS] = {}; 689d38ceaf9SAlex Deucher long r; 690d38ceaf9SAlex Deucher 691d38ceaf9SAlex Deucher seq[ring->idx] = atomic64_read(&ring->fence_drv.last_seq) + 1ULL; 692d38ceaf9SAlex Deucher if (seq[ring->idx] >= ring->fence_drv.sync_seq[ring->idx]) { 693d38ceaf9SAlex Deucher /* nothing to wait for, last_seq is 694d38ceaf9SAlex Deucher already the last emited fence */ 695d38ceaf9SAlex Deucher return -ENOENT; 696d38ceaf9SAlex Deucher } 697d38ceaf9SAlex Deucher r = amdgpu_fence_wait_seq_timeout(ring->adev, seq, false, MAX_SCHEDULE_TIMEOUT); 698d38ceaf9SAlex Deucher if (r < 0) 699d38ceaf9SAlex Deucher return r; 700d38ceaf9SAlex Deucher return 0; 701d38ceaf9SAlex Deucher } 702d38ceaf9SAlex Deucher 703d38ceaf9SAlex Deucher /** 704d38ceaf9SAlex Deucher * amdgpu_fence_wait_empty - wait for all fences to signal 705d38ceaf9SAlex Deucher * 706d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 707d38ceaf9SAlex Deucher * @ring: ring index the fence is associated with 708d38ceaf9SAlex Deucher * 709d38ceaf9SAlex Deucher * Wait for all fences on the requested ring to signal (all asics). 710d38ceaf9SAlex Deucher * Returns 0 if the fences have passed, error for all other cases. 711d38ceaf9SAlex Deucher * Caller must hold ring lock. 712d38ceaf9SAlex Deucher */ 713d38ceaf9SAlex Deucher int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) 714d38ceaf9SAlex Deucher { 715d38ceaf9SAlex Deucher struct amdgpu_device *adev = ring->adev; 716d38ceaf9SAlex Deucher uint64_t seq[AMDGPU_MAX_RINGS] = {}; 717d38ceaf9SAlex Deucher long r; 718d38ceaf9SAlex Deucher 719d38ceaf9SAlex Deucher seq[ring->idx] = ring->fence_drv.sync_seq[ring->idx]; 720d38ceaf9SAlex Deucher if (!seq[ring->idx]) 721d38ceaf9SAlex Deucher return 0; 722d38ceaf9SAlex Deucher 723d38ceaf9SAlex Deucher r = amdgpu_fence_wait_seq_timeout(adev, seq, false, MAX_SCHEDULE_TIMEOUT); 724d38ceaf9SAlex Deucher if (r < 0) { 725d38ceaf9SAlex Deucher if (r == -EDEADLK) 726d38ceaf9SAlex Deucher return -EDEADLK; 727d38ceaf9SAlex Deucher 728d38ceaf9SAlex Deucher dev_err(adev->dev, "error waiting for ring[%d] to become idle (%ld)\n", 729d38ceaf9SAlex Deucher ring->idx, r); 730d38ceaf9SAlex Deucher } 731d38ceaf9SAlex Deucher return 0; 732d38ceaf9SAlex Deucher } 733d38ceaf9SAlex Deucher 734d38ceaf9SAlex Deucher /** 735d38ceaf9SAlex Deucher * amdgpu_fence_ref - take a ref on a fence 736d38ceaf9SAlex Deucher * 737d38ceaf9SAlex Deucher * @fence: amdgpu fence object 738d38ceaf9SAlex Deucher * 739d38ceaf9SAlex Deucher * Take a reference on a fence (all asics). 740d38ceaf9SAlex Deucher * Returns the fence. 741d38ceaf9SAlex Deucher */ 742d38ceaf9SAlex Deucher struct amdgpu_fence *amdgpu_fence_ref(struct amdgpu_fence *fence) 743d38ceaf9SAlex Deucher { 744d38ceaf9SAlex Deucher fence_get(&fence->base); 745d38ceaf9SAlex Deucher return fence; 746d38ceaf9SAlex Deucher } 747d38ceaf9SAlex Deucher 748d38ceaf9SAlex Deucher /** 749d38ceaf9SAlex Deucher * amdgpu_fence_unref - remove a ref on a fence 750d38ceaf9SAlex Deucher * 751d38ceaf9SAlex Deucher * @fence: amdgpu fence object 752d38ceaf9SAlex Deucher * 753d38ceaf9SAlex Deucher * Remove a reference on a fence (all asics). 754d38ceaf9SAlex Deucher */ 755d38ceaf9SAlex Deucher void amdgpu_fence_unref(struct amdgpu_fence **fence) 756d38ceaf9SAlex Deucher { 757d38ceaf9SAlex Deucher struct amdgpu_fence *tmp = *fence; 758d38ceaf9SAlex Deucher 759d38ceaf9SAlex Deucher *fence = NULL; 760d38ceaf9SAlex Deucher if (tmp) 761d38ceaf9SAlex Deucher fence_put(&tmp->base); 762d38ceaf9SAlex Deucher } 763d38ceaf9SAlex Deucher 764d38ceaf9SAlex Deucher /** 765d38ceaf9SAlex Deucher * amdgpu_fence_count_emitted - get the count of emitted fences 766d38ceaf9SAlex Deucher * 767d38ceaf9SAlex Deucher * @ring: ring the fence is associated with 768d38ceaf9SAlex Deucher * 769d38ceaf9SAlex Deucher * Get the number of fences emitted on the requested ring (all asics). 770d38ceaf9SAlex Deucher * Returns the number of emitted fences on the ring. Used by the 771d38ceaf9SAlex Deucher * dynpm code to ring track activity. 772d38ceaf9SAlex Deucher */ 773d38ceaf9SAlex Deucher unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) 774d38ceaf9SAlex Deucher { 775d38ceaf9SAlex Deucher uint64_t emitted; 776d38ceaf9SAlex Deucher 777d38ceaf9SAlex Deucher /* We are not protected by ring lock when reading the last sequence 778d38ceaf9SAlex Deucher * but it's ok to report slightly wrong fence count here. 779d38ceaf9SAlex Deucher */ 780d38ceaf9SAlex Deucher amdgpu_fence_process(ring); 781d38ceaf9SAlex Deucher emitted = ring->fence_drv.sync_seq[ring->idx] 782d38ceaf9SAlex Deucher - atomic64_read(&ring->fence_drv.last_seq); 783d38ceaf9SAlex Deucher /* to avoid 32bits warp around */ 784d38ceaf9SAlex Deucher if (emitted > 0x10000000) 785d38ceaf9SAlex Deucher emitted = 0x10000000; 786d38ceaf9SAlex Deucher 787d38ceaf9SAlex Deucher return (unsigned)emitted; 788d38ceaf9SAlex Deucher } 789d38ceaf9SAlex Deucher 790d38ceaf9SAlex Deucher /** 791d38ceaf9SAlex Deucher * amdgpu_fence_need_sync - do we need a semaphore 792d38ceaf9SAlex Deucher * 793d38ceaf9SAlex Deucher * @fence: amdgpu fence object 794d38ceaf9SAlex Deucher * @dst_ring: which ring to check against 795d38ceaf9SAlex Deucher * 796d38ceaf9SAlex Deucher * Check if the fence needs to be synced against another ring 797d38ceaf9SAlex Deucher * (all asics). If so, we need to emit a semaphore. 798d38ceaf9SAlex Deucher * Returns true if we need to sync with another ring, false if 799d38ceaf9SAlex Deucher * not. 800d38ceaf9SAlex Deucher */ 801d38ceaf9SAlex Deucher bool amdgpu_fence_need_sync(struct amdgpu_fence *fence, 802d38ceaf9SAlex Deucher struct amdgpu_ring *dst_ring) 803d38ceaf9SAlex Deucher { 804d38ceaf9SAlex Deucher struct amdgpu_fence_driver *fdrv; 805d38ceaf9SAlex Deucher 806d38ceaf9SAlex Deucher if (!fence) 807d38ceaf9SAlex Deucher return false; 808d38ceaf9SAlex Deucher 809d38ceaf9SAlex Deucher if (fence->ring == dst_ring) 810d38ceaf9SAlex Deucher return false; 811d38ceaf9SAlex Deucher 812d38ceaf9SAlex Deucher /* we are protected by the ring mutex */ 813d38ceaf9SAlex Deucher fdrv = &dst_ring->fence_drv; 814d38ceaf9SAlex Deucher if (fence->seq <= fdrv->sync_seq[fence->ring->idx]) 815d38ceaf9SAlex Deucher return false; 816d38ceaf9SAlex Deucher 817d38ceaf9SAlex Deucher return true; 818d38ceaf9SAlex Deucher } 819d38ceaf9SAlex Deucher 820d38ceaf9SAlex Deucher /** 821d38ceaf9SAlex Deucher * amdgpu_fence_note_sync - record the sync point 822d38ceaf9SAlex Deucher * 823d38ceaf9SAlex Deucher * @fence: amdgpu fence object 824d38ceaf9SAlex Deucher * @dst_ring: which ring to check against 825d38ceaf9SAlex Deucher * 826d38ceaf9SAlex Deucher * Note the sequence number at which point the fence will 827d38ceaf9SAlex Deucher * be synced with the requested ring (all asics). 828d38ceaf9SAlex Deucher */ 829d38ceaf9SAlex Deucher void amdgpu_fence_note_sync(struct amdgpu_fence *fence, 830d38ceaf9SAlex Deucher struct amdgpu_ring *dst_ring) 831d38ceaf9SAlex Deucher { 832d38ceaf9SAlex Deucher struct amdgpu_fence_driver *dst, *src; 833d38ceaf9SAlex Deucher unsigned i; 834d38ceaf9SAlex Deucher 835d38ceaf9SAlex Deucher if (!fence) 836d38ceaf9SAlex Deucher return; 837d38ceaf9SAlex Deucher 838d38ceaf9SAlex Deucher if (fence->ring == dst_ring) 839d38ceaf9SAlex Deucher return; 840d38ceaf9SAlex Deucher 841d38ceaf9SAlex Deucher /* we are protected by the ring mutex */ 842d38ceaf9SAlex Deucher src = &fence->ring->fence_drv; 843d38ceaf9SAlex Deucher dst = &dst_ring->fence_drv; 844d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 845d38ceaf9SAlex Deucher if (i == dst_ring->idx) 846d38ceaf9SAlex Deucher continue; 847d38ceaf9SAlex Deucher 848d38ceaf9SAlex Deucher dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]); 849d38ceaf9SAlex Deucher } 850d38ceaf9SAlex Deucher } 851d38ceaf9SAlex Deucher 852d38ceaf9SAlex Deucher /** 853d38ceaf9SAlex Deucher * amdgpu_fence_driver_start_ring - make the fence driver 854d38ceaf9SAlex Deucher * ready for use on the requested ring. 855d38ceaf9SAlex Deucher * 856d38ceaf9SAlex Deucher * @ring: ring to start the fence driver on 857d38ceaf9SAlex Deucher * @irq_src: interrupt source to use for this ring 858d38ceaf9SAlex Deucher * @irq_type: interrupt type to use for this ring 859d38ceaf9SAlex Deucher * 860d38ceaf9SAlex Deucher * Make the fence driver ready for processing (all asics). 861d38ceaf9SAlex Deucher * Not all asics have all rings, so each asic will only 862d38ceaf9SAlex Deucher * start the fence driver on the rings it has. 863d38ceaf9SAlex Deucher * Returns 0 for success, errors for failure. 864d38ceaf9SAlex Deucher */ 865d38ceaf9SAlex Deucher int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, 866d38ceaf9SAlex Deucher struct amdgpu_irq_src *irq_src, 867d38ceaf9SAlex Deucher unsigned irq_type) 868d38ceaf9SAlex Deucher { 869d38ceaf9SAlex Deucher struct amdgpu_device *adev = ring->adev; 870d38ceaf9SAlex Deucher uint64_t index; 871d38ceaf9SAlex Deucher 872d38ceaf9SAlex Deucher if (ring != &adev->uvd.ring) { 873d38ceaf9SAlex Deucher ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs]; 874d38ceaf9SAlex Deucher ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4); 875d38ceaf9SAlex Deucher } else { 876d38ceaf9SAlex Deucher /* put fence directly behind firmware */ 877d38ceaf9SAlex Deucher index = ALIGN(adev->uvd.fw->size, 8); 878d38ceaf9SAlex Deucher ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index; 879d38ceaf9SAlex Deucher ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index; 880d38ceaf9SAlex Deucher } 881d38ceaf9SAlex Deucher amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq)); 882c6a4079bSChunming Zhou amdgpu_irq_get(adev, irq_src, irq_type); 883c6a4079bSChunming Zhou 884d38ceaf9SAlex Deucher ring->fence_drv.irq_src = irq_src; 885d38ceaf9SAlex Deucher ring->fence_drv.irq_type = irq_type; 886c6a4079bSChunming Zhou ring->fence_drv.initialized = true; 887c6a4079bSChunming Zhou 888d38ceaf9SAlex Deucher dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, " 889d38ceaf9SAlex Deucher "cpu addr 0x%p\n", ring->idx, 890d38ceaf9SAlex Deucher ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr); 891d38ceaf9SAlex Deucher return 0; 892d38ceaf9SAlex Deucher } 893d38ceaf9SAlex Deucher 894d38ceaf9SAlex Deucher /** 895d38ceaf9SAlex Deucher * amdgpu_fence_driver_init_ring - init the fence driver 896d38ceaf9SAlex Deucher * for the requested ring. 897d38ceaf9SAlex Deucher * 898d38ceaf9SAlex Deucher * @ring: ring to init the fence driver on 899d38ceaf9SAlex Deucher * 900d38ceaf9SAlex Deucher * Init the fence driver for the requested ring (all asics). 901d38ceaf9SAlex Deucher * Helper function for amdgpu_fence_driver_init(). 902d38ceaf9SAlex Deucher */ 903d38ceaf9SAlex Deucher void amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) 904d38ceaf9SAlex Deucher { 905d38ceaf9SAlex Deucher int i; 906d38ceaf9SAlex Deucher 907d38ceaf9SAlex Deucher ring->fence_drv.cpu_addr = NULL; 908d38ceaf9SAlex Deucher ring->fence_drv.gpu_addr = 0; 909d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) 910d38ceaf9SAlex Deucher ring->fence_drv.sync_seq[i] = 0; 911d38ceaf9SAlex Deucher 912d38ceaf9SAlex Deucher atomic64_set(&ring->fence_drv.last_seq, 0); 913d38ceaf9SAlex Deucher ring->fence_drv.initialized = false; 914d38ceaf9SAlex Deucher 915d38ceaf9SAlex Deucher INIT_DELAYED_WORK(&ring->fence_drv.lockup_work, 916d38ceaf9SAlex Deucher amdgpu_fence_check_lockup); 917d38ceaf9SAlex Deucher ring->fence_drv.ring = ring; 918b80d8475SAlex Deucher 919b80d8475SAlex Deucher if (amdgpu_enable_scheduler) { 920b80d8475SAlex Deucher ring->scheduler = amd_sched_create((void *)ring->adev, 921c1b69ed0SChunming Zhou &amdgpu_sched_ops, 9224afcb303SJammy Zhou ring->idx, 5, 0, 9234afcb303SJammy Zhou amdgpu_sched_hw_submission); 924b80d8475SAlex Deucher if (!ring->scheduler) 925b80d8475SAlex Deucher DRM_ERROR("Failed to create scheduler on ring %d.\n", 926b80d8475SAlex Deucher ring->idx); 927b80d8475SAlex Deucher } 928d38ceaf9SAlex Deucher } 929d38ceaf9SAlex Deucher 930d38ceaf9SAlex Deucher /** 931d38ceaf9SAlex Deucher * amdgpu_fence_driver_init - init the fence driver 932d38ceaf9SAlex Deucher * for all possible rings. 933d38ceaf9SAlex Deucher * 934d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 935d38ceaf9SAlex Deucher * 936d38ceaf9SAlex Deucher * Init the fence driver for all possible rings (all asics). 937d38ceaf9SAlex Deucher * Not all asics have all rings, so each asic will only 938d38ceaf9SAlex Deucher * start the fence driver on the rings it has using 939d38ceaf9SAlex Deucher * amdgpu_fence_driver_start_ring(). 940d38ceaf9SAlex Deucher * Returns 0 for success. 941d38ceaf9SAlex Deucher */ 942d38ceaf9SAlex Deucher int amdgpu_fence_driver_init(struct amdgpu_device *adev) 943d38ceaf9SAlex Deucher { 944d38ceaf9SAlex Deucher init_waitqueue_head(&adev->fence_queue); 945d38ceaf9SAlex Deucher if (amdgpu_debugfs_fence_init(adev)) 946d38ceaf9SAlex Deucher dev_err(adev->dev, "fence debugfs file creation failed\n"); 947d38ceaf9SAlex Deucher 948d38ceaf9SAlex Deucher return 0; 949d38ceaf9SAlex Deucher } 950d38ceaf9SAlex Deucher 951d38ceaf9SAlex Deucher /** 952d38ceaf9SAlex Deucher * amdgpu_fence_driver_fini - tear down the fence driver 953d38ceaf9SAlex Deucher * for all possible rings. 954d38ceaf9SAlex Deucher * 955d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 956d38ceaf9SAlex Deucher * 957d38ceaf9SAlex Deucher * Tear down the fence driver for all possible rings (all asics). 958d38ceaf9SAlex Deucher */ 959d38ceaf9SAlex Deucher void amdgpu_fence_driver_fini(struct amdgpu_device *adev) 960d38ceaf9SAlex Deucher { 961d38ceaf9SAlex Deucher int i, r; 962d38ceaf9SAlex Deucher 963d38ceaf9SAlex Deucher mutex_lock(&adev->ring_lock); 964d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 965d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 966d38ceaf9SAlex Deucher if (!ring || !ring->fence_drv.initialized) 967d38ceaf9SAlex Deucher continue; 968d38ceaf9SAlex Deucher r = amdgpu_fence_wait_empty(ring); 969d38ceaf9SAlex Deucher if (r) { 970d38ceaf9SAlex Deucher /* no need to trigger GPU reset as we are unloading */ 971d38ceaf9SAlex Deucher amdgpu_fence_driver_force_completion(adev); 972d38ceaf9SAlex Deucher } 973d38ceaf9SAlex Deucher wake_up_all(&adev->fence_queue); 974c6a4079bSChunming Zhou amdgpu_irq_put(adev, ring->fence_drv.irq_src, 975c6a4079bSChunming Zhou ring->fence_drv.irq_type); 976b80d8475SAlex Deucher if (ring->scheduler) 977b80d8475SAlex Deucher amd_sched_destroy(ring->scheduler); 978d38ceaf9SAlex Deucher ring->fence_drv.initialized = false; 979d38ceaf9SAlex Deucher } 980d38ceaf9SAlex Deucher mutex_unlock(&adev->ring_lock); 981d38ceaf9SAlex Deucher } 982d38ceaf9SAlex Deucher 983d38ceaf9SAlex Deucher /** 9845ceb54c6SAlex Deucher * amdgpu_fence_driver_suspend - suspend the fence driver 9855ceb54c6SAlex Deucher * for all possible rings. 9865ceb54c6SAlex Deucher * 9875ceb54c6SAlex Deucher * @adev: amdgpu device pointer 9885ceb54c6SAlex Deucher * 9895ceb54c6SAlex Deucher * Suspend the fence driver for all possible rings (all asics). 9905ceb54c6SAlex Deucher */ 9915ceb54c6SAlex Deucher void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) 9925ceb54c6SAlex Deucher { 9935ceb54c6SAlex Deucher int i, r; 9945ceb54c6SAlex Deucher 9955ceb54c6SAlex Deucher mutex_lock(&adev->ring_lock); 9965ceb54c6SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 9975ceb54c6SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 9985ceb54c6SAlex Deucher if (!ring || !ring->fence_drv.initialized) 9995ceb54c6SAlex Deucher continue; 10005ceb54c6SAlex Deucher 10015ceb54c6SAlex Deucher /* wait for gpu to finish processing current batch */ 10025ceb54c6SAlex Deucher r = amdgpu_fence_wait_empty(ring); 10035ceb54c6SAlex Deucher if (r) { 10045ceb54c6SAlex Deucher /* delay GPU reset to resume */ 10055ceb54c6SAlex Deucher amdgpu_fence_driver_force_completion(adev); 10065ceb54c6SAlex Deucher } 10075ceb54c6SAlex Deucher 10085ceb54c6SAlex Deucher /* disable the interrupt */ 10095ceb54c6SAlex Deucher amdgpu_irq_put(adev, ring->fence_drv.irq_src, 10105ceb54c6SAlex Deucher ring->fence_drv.irq_type); 10115ceb54c6SAlex Deucher } 10125ceb54c6SAlex Deucher mutex_unlock(&adev->ring_lock); 10135ceb54c6SAlex Deucher } 10145ceb54c6SAlex Deucher 10155ceb54c6SAlex Deucher /** 10165ceb54c6SAlex Deucher * amdgpu_fence_driver_resume - resume the fence driver 10175ceb54c6SAlex Deucher * for all possible rings. 10185ceb54c6SAlex Deucher * 10195ceb54c6SAlex Deucher * @adev: amdgpu device pointer 10205ceb54c6SAlex Deucher * 10215ceb54c6SAlex Deucher * Resume the fence driver for all possible rings (all asics). 10225ceb54c6SAlex Deucher * Not all asics have all rings, so each asic will only 10235ceb54c6SAlex Deucher * start the fence driver on the rings it has using 10245ceb54c6SAlex Deucher * amdgpu_fence_driver_start_ring(). 10255ceb54c6SAlex Deucher * Returns 0 for success. 10265ceb54c6SAlex Deucher */ 10275ceb54c6SAlex Deucher void amdgpu_fence_driver_resume(struct amdgpu_device *adev) 10285ceb54c6SAlex Deucher { 10295ceb54c6SAlex Deucher int i; 10305ceb54c6SAlex Deucher 10315ceb54c6SAlex Deucher mutex_lock(&adev->ring_lock); 10325ceb54c6SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 10335ceb54c6SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 10345ceb54c6SAlex Deucher if (!ring || !ring->fence_drv.initialized) 10355ceb54c6SAlex Deucher continue; 10365ceb54c6SAlex Deucher 10375ceb54c6SAlex Deucher /* enable the interrupt */ 10385ceb54c6SAlex Deucher amdgpu_irq_get(adev, ring->fence_drv.irq_src, 10395ceb54c6SAlex Deucher ring->fence_drv.irq_type); 10405ceb54c6SAlex Deucher } 10415ceb54c6SAlex Deucher mutex_unlock(&adev->ring_lock); 10425ceb54c6SAlex Deucher } 10435ceb54c6SAlex Deucher 10445ceb54c6SAlex Deucher /** 1045d38ceaf9SAlex Deucher * amdgpu_fence_driver_force_completion - force all fence waiter to complete 1046d38ceaf9SAlex Deucher * 1047d38ceaf9SAlex Deucher * @adev: amdgpu device pointer 1048d38ceaf9SAlex Deucher * 1049d38ceaf9SAlex Deucher * In case of GPU reset failure make sure no process keep waiting on fence 1050d38ceaf9SAlex Deucher * that will never complete. 1051d38ceaf9SAlex Deucher */ 1052d38ceaf9SAlex Deucher void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev) 1053d38ceaf9SAlex Deucher { 1054d38ceaf9SAlex Deucher int i; 1055d38ceaf9SAlex Deucher 1056d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; i++) { 1057d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 1058d38ceaf9SAlex Deucher if (!ring || !ring->fence_drv.initialized) 1059d38ceaf9SAlex Deucher continue; 1060d38ceaf9SAlex Deucher 1061d38ceaf9SAlex Deucher amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]); 1062d38ceaf9SAlex Deucher } 1063d38ceaf9SAlex Deucher } 1064d38ceaf9SAlex Deucher 1065d38ceaf9SAlex Deucher 1066d38ceaf9SAlex Deucher /* 1067d38ceaf9SAlex Deucher * Fence debugfs 1068d38ceaf9SAlex Deucher */ 1069d38ceaf9SAlex Deucher #if defined(CONFIG_DEBUG_FS) 1070d38ceaf9SAlex Deucher static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) 1071d38ceaf9SAlex Deucher { 1072d38ceaf9SAlex Deucher struct drm_info_node *node = (struct drm_info_node *)m->private; 1073d38ceaf9SAlex Deucher struct drm_device *dev = node->minor->dev; 1074d38ceaf9SAlex Deucher struct amdgpu_device *adev = dev->dev_private; 1075d38ceaf9SAlex Deucher int i, j; 1076d38ceaf9SAlex Deucher 1077d38ceaf9SAlex Deucher for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 1078d38ceaf9SAlex Deucher struct amdgpu_ring *ring = adev->rings[i]; 1079d38ceaf9SAlex Deucher if (!ring || !ring->fence_drv.initialized) 1080d38ceaf9SAlex Deucher continue; 1081d38ceaf9SAlex Deucher 1082d38ceaf9SAlex Deucher amdgpu_fence_process(ring); 1083d38ceaf9SAlex Deucher 1084344c19f9SChristian König seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name); 1085d38ceaf9SAlex Deucher seq_printf(m, "Last signaled fence 0x%016llx\n", 1086d38ceaf9SAlex Deucher (unsigned long long)atomic64_read(&ring->fence_drv.last_seq)); 1087d38ceaf9SAlex Deucher seq_printf(m, "Last emitted 0x%016llx\n", 1088d38ceaf9SAlex Deucher ring->fence_drv.sync_seq[i]); 1089d38ceaf9SAlex Deucher 1090d38ceaf9SAlex Deucher for (j = 0; j < AMDGPU_MAX_RINGS; ++j) { 1091d38ceaf9SAlex Deucher struct amdgpu_ring *other = adev->rings[j]; 1092344c19f9SChristian König if (i != j && other && other->fence_drv.initialized && 1093344c19f9SChristian König ring->fence_drv.sync_seq[j]) 1094d38ceaf9SAlex Deucher seq_printf(m, "Last sync to ring %d 0x%016llx\n", 1095d38ceaf9SAlex Deucher j, ring->fence_drv.sync_seq[j]); 1096d38ceaf9SAlex Deucher } 1097d38ceaf9SAlex Deucher } 1098d38ceaf9SAlex Deucher return 0; 1099d38ceaf9SAlex Deucher } 1100d38ceaf9SAlex Deucher 1101d38ceaf9SAlex Deucher static struct drm_info_list amdgpu_debugfs_fence_list[] = { 1102d38ceaf9SAlex Deucher {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, 1103d38ceaf9SAlex Deucher }; 1104d38ceaf9SAlex Deucher #endif 1105d38ceaf9SAlex Deucher 1106d38ceaf9SAlex Deucher int amdgpu_debugfs_fence_init(struct amdgpu_device *adev) 1107d38ceaf9SAlex Deucher { 1108d38ceaf9SAlex Deucher #if defined(CONFIG_DEBUG_FS) 1109d38ceaf9SAlex Deucher return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1); 1110d38ceaf9SAlex Deucher #else 1111d38ceaf9SAlex Deucher return 0; 1112d38ceaf9SAlex Deucher #endif 1113d38ceaf9SAlex Deucher } 1114d38ceaf9SAlex Deucher 1115d38ceaf9SAlex Deucher static const char *amdgpu_fence_get_driver_name(struct fence *fence) 1116d38ceaf9SAlex Deucher { 1117d38ceaf9SAlex Deucher return "amdgpu"; 1118d38ceaf9SAlex Deucher } 1119d38ceaf9SAlex Deucher 1120d38ceaf9SAlex Deucher static const char *amdgpu_fence_get_timeline_name(struct fence *f) 1121d38ceaf9SAlex Deucher { 1122d38ceaf9SAlex Deucher struct amdgpu_fence *fence = to_amdgpu_fence(f); 1123d38ceaf9SAlex Deucher return (const char *)fence->ring->name; 1124d38ceaf9SAlex Deucher } 1125d38ceaf9SAlex Deucher 1126d38ceaf9SAlex Deucher static inline bool amdgpu_test_signaled(struct amdgpu_fence *fence) 1127d38ceaf9SAlex Deucher { 1128d38ceaf9SAlex Deucher return test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags); 1129d38ceaf9SAlex Deucher } 1130d38ceaf9SAlex Deucher 1131d38ceaf9SAlex Deucher struct amdgpu_wait_cb { 1132d38ceaf9SAlex Deucher struct fence_cb base; 1133d38ceaf9SAlex Deucher struct task_struct *task; 1134d38ceaf9SAlex Deucher }; 1135d38ceaf9SAlex Deucher 1136d38ceaf9SAlex Deucher static void amdgpu_fence_wait_cb(struct fence *fence, struct fence_cb *cb) 1137d38ceaf9SAlex Deucher { 1138d38ceaf9SAlex Deucher struct amdgpu_wait_cb *wait = 1139d38ceaf9SAlex Deucher container_of(cb, struct amdgpu_wait_cb, base); 1140d38ceaf9SAlex Deucher wake_up_process(wait->task); 1141d38ceaf9SAlex Deucher } 1142d38ceaf9SAlex Deucher 1143d38ceaf9SAlex Deucher static signed long amdgpu_fence_default_wait(struct fence *f, bool intr, 1144d38ceaf9SAlex Deucher signed long t) 1145d38ceaf9SAlex Deucher { 1146d38ceaf9SAlex Deucher struct amdgpu_fence *fence = to_amdgpu_fence(f); 1147d38ceaf9SAlex Deucher struct amdgpu_device *adev = fence->ring->adev; 1148d38ceaf9SAlex Deucher struct amdgpu_wait_cb cb; 1149d38ceaf9SAlex Deucher 1150d38ceaf9SAlex Deucher cb.task = current; 1151d38ceaf9SAlex Deucher 1152d38ceaf9SAlex Deucher if (fence_add_callback(f, &cb.base, amdgpu_fence_wait_cb)) 1153d38ceaf9SAlex Deucher return t; 1154d38ceaf9SAlex Deucher 1155d38ceaf9SAlex Deucher while (t > 0) { 1156d38ceaf9SAlex Deucher if (intr) 1157d38ceaf9SAlex Deucher set_current_state(TASK_INTERRUPTIBLE); 1158d38ceaf9SAlex Deucher else 1159d38ceaf9SAlex Deucher set_current_state(TASK_UNINTERRUPTIBLE); 1160d38ceaf9SAlex Deucher 1161d38ceaf9SAlex Deucher /* 1162d38ceaf9SAlex Deucher * amdgpu_test_signaled must be called after 1163d38ceaf9SAlex Deucher * set_current_state to prevent a race with wake_up_process 1164d38ceaf9SAlex Deucher */ 1165d38ceaf9SAlex Deucher if (amdgpu_test_signaled(fence)) 1166d38ceaf9SAlex Deucher break; 1167d38ceaf9SAlex Deucher 1168d38ceaf9SAlex Deucher if (adev->needs_reset) { 1169d38ceaf9SAlex Deucher t = -EDEADLK; 1170d38ceaf9SAlex Deucher break; 1171d38ceaf9SAlex Deucher } 1172d38ceaf9SAlex Deucher 1173d38ceaf9SAlex Deucher t = schedule_timeout(t); 1174d38ceaf9SAlex Deucher 1175d38ceaf9SAlex Deucher if (t > 0 && intr && signal_pending(current)) 1176d38ceaf9SAlex Deucher t = -ERESTARTSYS; 1177d38ceaf9SAlex Deucher } 1178d38ceaf9SAlex Deucher 1179d38ceaf9SAlex Deucher __set_current_state(TASK_RUNNING); 1180d38ceaf9SAlex Deucher fence_remove_callback(f, &cb.base); 1181d38ceaf9SAlex Deucher 1182d38ceaf9SAlex Deucher return t; 1183d38ceaf9SAlex Deucher } 1184d38ceaf9SAlex Deucher 1185d38ceaf9SAlex Deucher const struct fence_ops amdgpu_fence_ops = { 1186d38ceaf9SAlex Deucher .get_driver_name = amdgpu_fence_get_driver_name, 1187d38ceaf9SAlex Deucher .get_timeline_name = amdgpu_fence_get_timeline_name, 1188d38ceaf9SAlex Deucher .enable_signaling = amdgpu_fence_enable_signaling, 1189d38ceaf9SAlex Deucher .signaled = amdgpu_fence_is_signaled, 1190d38ceaf9SAlex Deucher .wait = amdgpu_fence_default_wait, 1191d38ceaf9SAlex Deucher .release = NULL, 1192d38ceaf9SAlex Deucher }; 1193